매스 MATH

레오의 AI/코딩/트레이딩/아비트라지1d ago

GPT-6 Astra improves Qwen performance by directly reading and editing neural network weights - In early experiments where GPT-6 Astra directly reads and edits neural network weights or writes them from scratch without training loops or gradient descent, it showed ability to identify model behavior and fix errors - Modified 0.016% of Qwen3.5-0.8B's weights to reduce repetitive reasoning, with MATH-500 improving from 7.9% to 48.1% and IFEval from 21.0% to 62.3% at the same sampling settings - In small neural networks, identified games and languages learned from weights alone without running the model, and even accurately found attention heads responsible for text copying in language models with input/output layers removed - Environmental execution or feedback… https://news.hada.io/topic?id=36478

289Spread to 2 channelsShow Korean original